Information recommendation model training method, information recommendation method, device and server

Abstract

The present disclosure relates to a training method, information recommendation method and device for an information recommendation model. The information recommendation model training method includes: obtaining initial sample features and actual operation results of sample accounts for sample recommendation information; the initial sample features are obtained based on sample account features and sample information features; the initial sample features are input into the information recommendation model to be trained to obtain predicted operation results of the sample account for sample recommendation information; the model includes a gating structure and multiple feature extraction layers, the multiple feature extraction layers are used to output multiple target sample features, the output value of the gating structure represents the importance of each target sample feature, and the predicted operation result is obtained based on the multiple target sample features and the output value of the gating structure; based on the difference between the predicted operation result and the actual operation result, the information recommendation model is trained to obtain a trained information recommendation model. The present disclosure can improve the recommendation accuracy of the trained information recommendation model.

Description

Technical Field

[0001] The present disclosure relates to the technical field of information recommendation, and in particular to a training method for an information recommendation model, an information recommendation method, a device, an electronic device, and a storage medium. Background Art

[0002] With the development of information recommendation technology, a technology for recommending information using information recommendation models has emerged. By training the information recommendation model, when making information recommendations, the information features of the recommended information contained in the recommendation information library and the user features of the recommending user who recommends the information can be extracted, and the information features and user features are input into the recommendation model. The model can then output the probability of the user accessing each recommended information, thereby providing users with recommended information with a high access probability.

[0003] In the related technology, during the training process of the current information recommendation model, after collecting samples, different levels of networks can be used to extract shallow features and deep features of the samples respectively, and the shallow features and the deep features are fused by direct addition to realize the training of the information recommendation model. However, this direct addition fusion method is relatively simple, so the recommendation accuracy of the trained information recommendation model is low. Summary of the invention

[0004] The present disclosure provides a training method for an information recommendation model, an information recommendation method, an apparatus, an electronic device, and a storage medium, so as to at least solve the problem of low recommendation accuracy of the information recommendation model in the related art. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a method for training an information recommendation model is provided, comprising:

[0006] Acquire initial sample features and actual operation results of the sample account for the sample recommendation information; the initial sample features are obtained based on the sample account features of the sample account and the sample information features of the sample recommendation information;

[0007] Input the initial sample features into the information recommendation model to be trained, and obtain the prediction operation result of the sample account for the sample recommendation information; the information recommendation model includes a gating structure and multiple feature extraction layers, the multiple feature extraction layers are used to output multiple target sample features for the initial sample features, the output value of the gating structure is obtained according to the initial sample features, and is used to characterize the importance of each target sample feature, and the prediction operation result is obtained according to the multiple target sample features and the output value of the gating structure;

[0008] Based on the difference between the predicted operation result and the actual operation result, the information recommendation model to be trained is trained to obtain a trained information recommendation model.

[0009] In an exemplary embodiment, the inputting of the initial sample features into the information recommendation model to be trained to obtain the prediction operation result of the sample account for the sample recommendation information includes: inputting the initial sample features into the information recommendation model to be trained, and obtaining multiple target sample features for the initial sample features through multiple feature extraction layers of the information recommendation model; inputting the initial sample features into the gating structure, and obtaining the output value of the gating structure corresponding to each target sample feature through the gating structure; performing weighted processing on each target sample feature according to the output value of the gating structure, and obtaining the prediction operation result using the result of the weighted processing.

[0010] In an exemplary embodiment, the multiple feature extraction layers of the recommendation model are used to obtain multiple target sample features for the initial sample features, including: obtaining a current feature extraction layer from the multiple feature extraction layers, and obtaining a current input sample feature for inputting the current feature extraction layer; inputting the current input sample feature into the current feature extraction layer, and obtaining a target sample feature corresponding to the current feature extraction layer through feature extraction layer parameters corresponding to the current feature extraction layer.

[0011] In an exemplary embodiment, the obtaining of current input sample features for inputting the current feature extraction layer includes: when the current feature extraction layer is the first feature extraction layer of the information recommendation model to be trained, using the initial sample features as the current input sample features; when the current feature extraction layer is not the first feature extraction layer of the information recommendation model to be trained, using the target sample features corresponding to the previous feature extraction layer as the current input sample features.

[0012] In an exemplary embodiment, the inputting the initial sample features into the gating structure and obtaining the output values ​​of the gating structure corresponding to each target sample feature respectively through the gating structure includes: obtaining the gated input sample features from the initial sample features; inputting the gated input sample features into the gating structure and obtaining the output values ​​of the gating structure through the gating structure.

[0013] In an exemplary embodiment, the initial sample feature is composed of a first initial sample feature and a second initial sample feature; the first initial sample feature is an initial sample feature used to identify the sample account, or used to identify the sample recommendation information, and the second initial sample feature is an initial sample feature other than the first initial sample feature among the initial sample features; obtaining the gated input sample feature from the initial sample features includes: filtering out the second gated input sample feature from the second initial sample feature according to preset gated input filtering information; using the first initial sample feature as the first gated input sample feature, and using the first gated input sample feature and the second gated input sample feature as the gated input sample features.

[0014] According to a second aspect of an embodiment of the present disclosure, there is provided an information recommendation method, comprising:

[0015] Acquire information recommendation features; wherein the information recommendation features are obtained based on account features of the target account and information features of the candidate recommended information;

[0016] Inputting the information recommendation feature into a trained information recommendation model to obtain a predicted operation result of the target account for the candidate recommended information; wherein the information recommendation model is obtained by the information recommendation model training method as described in any one of the embodiments of the first aspect;

[0017] According to the prediction operation result, target recommendation information for the target account is screened out from the candidate recommendation information, and the target recommendation information is recommended to the target account.

[0018] According to a third aspect of an embodiment of the present disclosure, there is provided a training device for an information recommendation model, comprising:

[0019] A sample data acquisition unit is configured to acquire initial sample features and actual operation results of a sample account for sample recommendation information; the initial sample features are acquired based on sample account features of the sample account and sample information features of the sample recommendation information;

[0020] A prediction result acquisition unit is configured to execute inputting the initial sample feature into the information recommendation model to be trained, and obtain the prediction operation result of the sample account for the sample recommendation information; the information recommendation model includes a gating structure and a plurality of feature extraction layers, the plurality of feature extraction layers are used to output a plurality of target sample features for the initial sample feature, the output value of the gating structure is obtained according to the initial sample feature, and is used to characterize the importance of each target sample feature, and the prediction operation result is obtained according to the plurality of target sample features and the output value of the gating structure;

[0021] The recommendation model training unit is configured to perform training on the information recommendation model to be trained based on the difference between the predicted operation result and the actual operation result to obtain a trained information recommendation model.

[0022] In an exemplary embodiment, the prediction result acquisition unit is further configured to execute inputting the initial sample features into the information recommendation model to be trained, and obtaining multiple target sample features for the initial sample features through multiple feature extraction layers of the information recommendation model; inputting the initial sample features into the gating structure, and obtaining the output values ​​of the gating structure corresponding to each target sample feature through the gating structure; performing weighted processing on each target sample feature according to the output value of the gating structure, and obtaining the prediction operation result using the result of the weighted processing.

[0023] In an exemplary embodiment, the prediction result acquisition unit is further configured to execute acquiring the current feature extraction layer from the multiple feature extraction layers, and acquiring the current input sample features for inputting the current feature extraction layer; inputting the current input sample features into the current feature extraction layer, and obtaining the target sample features corresponding to the current feature extraction layer through the feature extraction layer parameters corresponding to the current feature extraction layer.

[0024] In an exemplary embodiment, the prediction result acquisition unit is further configured to use the initial sample features as the current input sample features when the current feature extraction layer is the first feature extraction layer of the information recommendation model to be trained; and to use the target sample features corresponding to the previous feature extraction layer as the current input sample features when the current feature extraction layer is not the first feature extraction layer of the information recommendation model to be trained.

[0025] In an exemplary embodiment, the prediction result acquisition unit is further configured to acquire gated input sample features from the initial sample features; input the gated input sample features into the gating structure, and obtain the output value of the gating structure through the gating structure.

[0026] In an exemplary embodiment, the initial sample feature is composed of a first initial sample feature and a second initial sample feature; the first initial sample feature is an initial sample feature used to identify the sample account, or used to identify the sample recommendation information, and the second initial sample feature is an initial sample feature other than the first initial sample feature among the initial sample features; the prediction result acquisition unit is further configured to perform a filter according to preset gated input filtering information, and filter out a second gated input sample feature from the second initial sample feature; use the first initial sample feature as the first gated input sample feature, and use the first gated input sample feature and the second gated input sample feature as the gated input sample features.

[0027] According to a fourth aspect of an embodiment of the present disclosure, there is provided an information recommendation device, including:

[0028] A recommendation feature acquisition unit is configured to acquire information recommendation features; wherein the information recommendation features are obtained according to account features of the target account and information features of the candidate recommended information;

[0029] a model prediction acquisition unit, configured to input the information recommendation feature into a trained information recommendation model to obtain a prediction operation result of the target account for the candidate recommended information; wherein the information recommendation model is obtained by the information recommendation model training method as described in any one of the embodiments of the first aspect;

[0030] The target information recommendation unit is configured to perform, according to the prediction operation result, filtering out target recommendation information for the target account from the candidate recommendation information, and recommending the target recommendation information to the target account.

[0031] According to a fifth aspect of an embodiment of the present disclosure, a server is provided, comprising: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the information recommendation model training method as described in any one of the embodiments of the first aspect, or the information recommendation method as described in any one of the embodiments of the second aspect.

[0032] According to a sixth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of a server, the server is enabled to execute the training method of the information recommendation model as described in any one of the embodiments in the first aspect, or the information recommendation method as described in any one of the embodiments in the second aspect.

[0033] According to a seventh aspect of an embodiment of the present disclosure, a computer program product is provided, wherein the computer program product includes instructions, and when the instructions are executed by a processor of a server, the server is able to execute the training method of the information recommendation model as described in any one of the embodiments of the first aspect, or the information recommendation method as described in any one of the embodiments of the second aspect.

[0034] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0035] By obtaining initial sample features and actual operation results of sample accounts for sample recommendation information; the initial sample features are obtained based on the sample account features of the sample account and the sample information features of the sample recommendation information; the initial sample features are input into the information recommendation model to be trained to obtain the predicted operation results of the sample account for the sample recommendation information; the information recommendation model includes a gating structure and multiple feature extraction layers, the multiple feature extraction layers are used to output multiple target sample features for the initial sample features, the output value of the gating structure is obtained based on the initial sample features, and is used to characterize the importance of each target sample feature, and the predicted operation result is obtained based on the multiple target sample features and the output value of the gating structure; based on the difference between the predicted operation result and the actual operation result, the information recommendation model to be trained is trained to obtain the trained information recommendation model. The present disclosure can obtain the importance of target sample features output by each feature extraction layer through the output value of the gating structure. At the same time, the prediction operation result output by the information recommendation model can be determined based on the target sample features and the output value of the gating structure, thereby realizing the training of the information recommendation model. Compared with the traditional technology of realizing the information recommendation model training by fusing the features output by each feature extraction layer by directly adding them, the present disclosure can realize the fusion of features based on the importance of the features output by each feature extraction layer, and the above-mentioned importance can be determined by the initial sample features input to the information recommendation model, realizing the adaptive fusion of the features output by each feature extraction layer, thereby improving the recommendation accuracy of the trained information recommendation model.

[0036] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0038] Figure 1 The figure is a flowchart of a method for training an information recommendation model according to an exemplary embodiment.

[0039] Figure 2 The figure is a schematic diagram of the structure of an information recommendation model according to an exemplary embodiment.

[0040] Figure 3 The figure is a flow chart of obtaining a prediction operation result according to an exemplary embodiment.

[0041] Figure 4 The flowchart of obtaining the output value of the gating structure is shown according to an exemplary embodiment.

[0042] Figure 5 The figure is a flowchart of an information recommendation method according to an exemplary embodiment.

[0043] Figure 6 It is a block diagram of a training device for an information recommendation model according to an exemplary embodiment.

[0044] Figure 7 The invention is a block diagram of an information recommendation device according to an exemplary embodiment.

[0045] Figure 8 It is a block diagram of a server according to an exemplary embodiment. DETAILED DESCRIPTION

[0046] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0047] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0048] It should also be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0049] Figure 1 is a flowchart of a training method for an information recommendation model according to an exemplary embodiment. Figure 1As shown, the training method of the information recommendation model is used in the server and includes the following steps.

[0050] In step S101, initial sample characteristics and actual operation results of the sample account for sample recommendation information are obtained; the initial sample characteristics are obtained based on the sample account characteristics of the sample account and the sample information characteristics of the sample recommendation information.

[0051] Among them, the information recommendation model is a neural network model used to recommend information to an account. For example, it can be a network model for recommending videos to an account. The model can input the account characteristics of the account and the information characteristics of the recommended information to be recommended, such as the video characteristics of the video, into the recommendation model. The recommendation model can determine the probability of the account performing an operation on the recommended information based on the account characteristics and information characteristics. For example, it can be the probability of the account accessing the recommended video or the first layer, etc., and then recommend recommended information that is more relevant to the user to the user based on the account's operation probability.

[0052] The sample data is the data used to train the above-mentioned information recommendation model. The sample data may include initial sample features and actual operation results of the sample account used to train the recommendation model for the sample recommendation information. The actual operation results may be used to characterize whether the sample account has performed operations such as clicks, likes, and shares on the sample recommendation information. The initial sample features may be composed of features of the sample account data, i.e., sample account features, and features of the sample recommendation information data, i.e., sample information features.

[0053] Specifically, when the server trains the information recommendation model, it can pre-collect data features of the training samples, such as sample account features of the sample accounts, sample information features of the sample recommendation information, and the actual access results of the above sample accounts for the above sample recommendation information, that is, whether the sample account accesses the above sample recommendation information. Afterwards, the initial sample features can be formed based on the collected sample account features and sample information features, and the initial sample features and the above actual operation results are used as the initial sample features for training the information recommendation model. For example, the sample account features and the sample information features can be represented in the form of vectors. After obtaining the vector form of the above features, the above features can be fused in series to form the initial sample features.

[0054] In step S102, the initial sample features are input into the information recommendation model to be trained to obtain the predicted operation results of the sample account for the sample recommendation information; the recommendation model includes a gating structure and multiple feature extraction layers, the multiple feature extraction layers are used to output multiple target sample features for the initial sample features, the output value of the gating structure is obtained according to the initial sample features, and is used to characterize the importance of each target sample feature, and the predicted operation results are obtained according to the multiple target sample features and the output value of the gating structure.

[0055] The output value of the gating structure refers to the output parameter of the gating structure carried in the information recommendation model. The output value can be determined based on the initial sample features input to the information recommendation model, and is used to characterize the importance of each target sample feature. The number of output values ​​can match the number of layers corresponding to the target sample features. The target sample features refer to the sample features obtained by performing feature processing of different dimensions on the initial sample features using different feature extraction layers of the information recommendation model. For example, the number of feature extraction layers used to process the initial sample feature A in the information recommendation model is 4. After the initial sample feature A is input into the information recommendation model, the target sample features of 4 different levels for the initial sample feature A can be output through the above 4 feature extraction layers, which can be target sample feature O1, target sample feature O2, target sample feature O3, and target sample feature O4. At the same time, the number of output values ​​of the gating structure can also be 4, which correspond to the above-mentioned target sample feature O1, target sample feature O2, target sample feature O3, and target sample feature O4 respectively, and are used to characterize the importance between the above-mentioned target sample feature O1, target sample feature O2, target sample feature O3, and target sample feature O4.

[0056] The predicted operation result refers to the predicted result output by the recommendation model, which can be the predicted operation probability of the sample account for the sample recommendation information. After the server inputs the initial sample features into the information recommendation model, it can output multiple target sample features for the initial sample features through the multiple feature extraction layers contained in the information recommendation model, and can obtain the output value used to characterize the importance of each of the above target sample features based on the initial sample features and the gating structure carried in the information recommendation model. Finally, the predicted operation result of the sample account for the sample recommendation information can be determined based on the above multiple target sample features and the output value of the gating structure.

[0057] like Figure 2 As shown, Figure 2It can be a schematic diagram of the model structure of an information recommendation model. It can be seen that the initial sample features input to the information recommendation model can be composed of vector 1, vector 2, ..., vector n. The information recommendation model can include multiple feature extraction layers, feature extraction layer 1, feature extraction layer 2, feature extraction layer 3, ..., feature extraction layer n. After the server inputs the initial sample features into the information recommendation model, multiple target sample features for the initial sample features can be obtained through the multiple feature extraction layers included in the information recommendation model, namely feature extraction layer 1, feature extraction layer 2, feature extraction layer 3, ..., feature extraction layer n, namely target sample feature 1, target sample feature 2, target sample feature 3, ..., target sample feature n. In addition, according to the input initial sample features and the gating structure included in the information recommendation model, the gated structure output values ​​corresponding to the above multiple target sample features can be obtained, so that the information recommendation model can obtain the final output result, namely the prediction operation result, through the above multiple target sample features and the gated structure output value used to characterize the importance of each target sample feature.

[0058] In step S103, based on the difference between the predicted operation result and the actual operation result, the information recommendation model to be trained is trained to obtain a trained information recommendation model.

[0059] After the server obtains the predicted operation result of the sample account for the sample recommendation information in step S102, it can also calculate the difference between the predicted operation result and the actual operation result of the sample account for the sample recommendation information. For example, a loss value between the predicted operation result and the actual operation result can be obtained through a pre-constructed loss function, and the loss value is used as the difference between the predicted operation result and the actual operation result. The information recommendation model is trained using the loss value to obtain a trained information recommendation model. For example, when the loss value of the information recommendation model is less than a loss value threshold, or when the number of training steps reaches a certain number, the training is stopped to obtain the final information recommendation model.

[0060] In the training method of the above-mentioned information recommendation model, by obtaining initial sample features and actual operation results of the sample account for sample recommendation information; the initial sample features are obtained based on the sample account features of the sample account and the sample information features of the sample recommendation information; the initial sample features are input into the information recommendation model to be trained to obtain the predicted operation results of the sample account for the sample recommendation information; the information recommendation model includes a gating structure and multiple feature extraction layers, the multiple feature extraction layers are used to output multiple target sample features for the initial sample features, the output value of the gating structure is obtained based on the initial sample features, and is used to characterize the importance of each target sample feature, and the predicted operation result is obtained based on the multiple target sample features and the output value of the gating structure; based on the difference between the predicted operation result and the actual operation result, the information recommendation model to be trained is trained to obtain a trained information recommendation model. The present disclosure can obtain the importance of target sample features output by each feature extraction layer through the output value of the gating structure. At the same time, the prediction operation result output by the information recommendation model can be determined based on the target sample features and the output value of the gating structure, thereby realizing the training of the information recommendation model. Compared with the traditional technology of realizing the information recommendation model training by fusing the features output by each feature extraction layer by directly adding them, the present disclosure can realize the fusion of features based on the importance of the features output by each feature extraction layer, and the above-mentioned importance can be determined by the initial sample features input to the information recommendation model, realizing the adaptive fusion of the features output by each feature extraction layer, thereby improving the recommendation accuracy of the trained information recommendation model.

[0061] In an exemplary embodiment, if Figure 3 As shown, step S102 may further include:

[0062] In step S301, the initial sample features are input into the information recommendation model to be trained, and multiple target sample features corresponding to the initial sample features are obtained through multiple feature extraction layers of the information recommendation model.

[0063] In this embodiment, multiple target sample features can be obtained by outputting multiple feature extraction layers respectively. The server can input the initial sample features into the information recommendation model including multiple feature extraction layers, and each feature extraction layer of the information recommendation model can output multiple target sample features for the initial sample features respectively. For example, the number of multiple feature extraction layers can be K, namely feature extraction layer 1, feature extraction layer 2, ..., feature extraction layer K. After the server inputs the initial sample features into the information recommendation model, it can obtain the corresponding multiple target sample features respectively through each feature extraction layer of the information recommendation model, and obtain target sample features O1, target sample features O2, ..., target sample features O3, ..., respectively. K .

[0064] In step S302, the initial sample features are input into the gating structure, and the output values ​​of the gating structure corresponding to each target sample feature are obtained through the gating structure.

[0065] The output value of the gating structure can be obtained by the server inputting the initial sample features into the gating structure of the information recommendation model, and the output value of the gating structure can be adapted to the target sample features output by each feature extraction layer. Specifically, when the server inputs the initial sample features into the information recommendation model, it can also input the initial sample features into the gating structure of the information recommendation model, and the gating structure outputs the output values ​​of the corresponding target sample features.

[0066] For example, for an information recommendation model including K feature extraction layers, the target sample features output by each feature extraction layer may be target sample feature O1, target sample feature O2, ..., target sample feature O K , then the number of output values ​​obtained by the gate structure can also be K, which are gate structure output value g1, gate structure output value g2, ..., gate structure output value g K , and the above gating structure adopts the MLP neural network structure, so the output values ​​of K gating structures can be obtained by the following formula:

[0067] [g1,g2,…,g k ]=F g (A,W g )=σ(A*W g )

[0068] Among them, A represents the initial sample features input to the gated structure, and W g It represents the structural parameters corresponding to the MLP neural network structure used for the gating structure, and σ represents the sigmoid activation function, that is, the S-shaped double bending function.

[0069] In step S303, weighted processing is performed on each target sample feature according to the output value of the gate structure, and the prediction operation result is obtained using the result of the weighted processing.

[0070] Finally, after obtaining the above-mentioned multiple target sample features, the output value of the gating structure corresponding to each target sample feature can be used to perform weighted processing on each target sample feature, and the result obtained after weighted processing can be used to finally obtain the prediction operation result output by the information recommendation model.

[0071] For example, the multiple target sample features may be target sample feature O1, target sample feature O2, ..., target sample feature O K, the output values ​​of each gate control structure can be gate control structure output value g1, gate control structure output value g2, ..., gate control structure output value g K , the server can perform weighted processing using the following formula:

[0072]

[0073] Among them, O * represents the result after weighted processing, O k represents the kth target sample feature, and g k It represents the output value of the gating parameter corresponding to the kth target sample feature.

[0074] After that, the final estimated value can be obtained through the final MLP neural network, which is achieved through the following calculation formula:

[0075] pred=F(O * *W * )=σ(O * *W)

[0076] Where σ is the sigmoid activation function, that is, the S-shaped double bending function, and pred represents the final prediction operation result.

[0077] In this embodiment, after the initial sample features are input into the information recommendation model, multiple target sample features can be obtained respectively through multiple feature extraction layers of the information recommendation model. The initial sample features can also be input into the gating structure of the information recommendation model, and the gating structure outputs the corresponding output value, and the prediction operation result is obtained in a weighted manner. This embodiment can characterize the importance of each target sample feature in the form of weights, thereby further improving the accuracy of the obtained prediction operation results.

[0078] Furthermore, step S301 may further include: obtaining a current feature extraction layer from multiple feature extraction layers, and obtaining current input sample features used to input the current feature extraction layer; inputting the current input sample features into the current feature extraction layer, and obtaining target sample features corresponding to the current feature extraction layer through feature extraction layer parameters corresponding to the current feature extraction layer.

[0079] The current feature extraction layer may refer to any one of the multiple feature extraction layers included in the information recommendation model, and the current input sample feature refers to the sample feature input to the current feature extraction layer. For example, if the current feature extraction layer is feature extraction layer 1, then the sample feature input to feature extraction layer 1 may be used as the current input sample feature. If the current feature extraction layer is feature extraction layer 2, then the sample feature input to feature extraction layer 2 may be used as the current input sample feature. In this embodiment, the server may obtain the current input sample feature corresponding to each current feature extraction layer, and by inputting the current input sample feature into the corresponding current feature extraction layer, the target sample feature corresponding to each current feature extraction layer may be output through the feature extraction layer parameters corresponding to the current feature extraction layer.

[0080] In this embodiment, the server can also obtain the input sample features corresponding to each feature extraction layer, and use the feature extraction layer parameters of each feature extraction layer to process the input sample features of each feature extraction layer, thereby outputting the target sample features of each feature extraction layer. In the above manner, the server can obtain the target sample features of each feature extraction layer, thereby improving the accuracy of the target sample features output by each feature extraction layer.

[0081] Furthermore, obtaining current input sample features for inputting the current feature extraction layer may further include: when the current feature extraction layer is the first feature extraction layer of the information recommendation model to be trained, using the initial sample features as the current input sample features; when the current feature extraction layer is not the first feature extraction layer of the information recommendation model to be trained, using the target sample features corresponding to the previous feature extraction layer as the current input sample features.

[0082] In this embodiment, there are multiple feature extraction layers, and the first feature extraction layer refers to the first feature extraction layer corresponding to the initial sample feature input to the information recommendation model. For example, the order of the multiple feature extraction layers is feature extraction layer 1, feature extraction layer 2, and feature extraction layer 3, then feature extraction layer 1 can be used as the first feature extraction layer in the information recommendation model to be trained. Specifically, if the current feature extraction layer is the first feature extraction layer of the information recommendation model, that is, feature extraction layer 1, then the server can use the initial sample feature as the current input sample feature to output the corresponding target sample feature 1, and if the current feature extraction layer is not the first feature extraction layer of the information recommendation model, then the server can use the previous feature extraction layer of the current feature extraction layer, that is, the target sample feature of the previous feature extraction layer, as the current input sample feature. For example, if the current feature extraction layer is feature extraction layer 2, which is not the first feature extraction layer of the information recommendation model, the server can use the target sample feature corresponding to the previous feature extraction layer of the current feature extraction layer, that is, the target sample feature 1 corresponding to feature extraction layer 1, as the current input sample feature, thereby obtaining the corresponding target sample feature 2 through feature extraction layer 2. If the current feature extraction layer is feature extraction layer 3, the server can use the target sample feature 2 corresponding to feature extraction layer 2 as the current input sample feature.

[0083] For example, the target sample features corresponding to each feature extraction layer can be represented by the following calculation formula:

[0084] O1=F1(A*W1)

[0085] O k =F k (O k-1 *W k )

[0086] The function F represents a single-layer feature extraction layer of the neural network, F(X,W) = σ(X*W), F k represents the feature extraction layer of the k-th neural network, W k represents the feature extraction layer parameters of the kth feature extraction layer, σ represents the ReLU activation function, i.e., the linear rectification function, O k It represents the target sample features output by the kth feature extraction layer. It can be seen that for the first feature extraction layer, its input is the initial sample feature A, and for other feature extraction layers, its input can be the target sample features output by the k-1th feature extraction layer. Through the above method, target sample features of various depths can be obtained, which represents the information extracted by the model at different depths.

[0087] In this embodiment, if the current feature extraction layer is the first feature extraction layer, the initial sample feature can be used as the current input sample feature. If the current feature extraction layer is not the first feature extraction layer, the target sample feature of the previous feature extraction layer can be used as the current input sample feature. In the above manner, the target sample feature of each feature extraction layer can be obtained, thereby obtaining target sample features of different depths.

[0088] In addition, in an exemplary embodiment, Figure 4 As shown, step S302 may further include:

[0089] In step S401, gated input sample features are obtained from initial sample features.

[0090] The gated input sample features refer to the initial sample features used to input the gated structure in the initial sample features. In this embodiment, in order to improve the operational efficiency of the information recommendation model, when the initial sample features are input into the gated structure to obtain the output value of the gated structure, not all the initial sample features are input into the gated structure to obtain the output value of the above-mentioned gated structure, but first from the initial sample features, according to the set feature selection rules, a part of the sample features as the input to the gated structure, that is, the gated input sample features, are screened out. For example, the initial sample features can be composed of 10 feature vectors, then before they are input into the gated structure, the above-mentioned 10 feature vectors can be preliminarily screened, for example, 5 of the feature vectors are screened out as the gated input sample features for input into the gated structure.

[0091] In step S402, the gated input sample feature is input into the gated structure, and the output value of the gated structure is obtained through the gated structure.

[0092] After obtaining the gated input sample features in step S401, the gated input sample features may be further input into the gating structure of the information recommendation model, so that an output value of the gating structure adapted to the gated input sample features may be obtained through the gating structure.

[0093] In this embodiment, when using the gated structure to obtain the output value of the gated structure, the gated input sample features can be first screened out from the initial sample features, and the gated input sample features can be input into the gated structure to obtain the output value of the above-mentioned gated structure. Compared with inputting all the initial sample features into the gated structure to obtain the above-mentioned output value, this embodiment can reduce the amount of calculation of the gated structure to obtain the output value, thereby further improving the calculation efficiency of the information recommendation model.

[0094] Further, the initial sample feature is composed of a first initial sample feature and a second initial sample feature; the first initial sample feature is an initial sample feature used to identify a sample account or to identify sample recommendation information, and the second initial sample feature is an initial sample feature other than the first initial sample feature among the initial sample features; step S401 may further include: according to preset gated input filtering information, filtering out a second gated input sample feature from the second initial sample feature; using the first initial sample feature as the first gated input sample feature, and using the first gated input sample feature and the second gated input sample feature as gated input sample features.

[0095] The first initial sample feature refers to the initial sample feature used to identify the sample account or the sample recommendation information, for example, the account ID feature of the sample account, that is, the account information number feature, or the recommendation information ID feature of the sample recommendation information, that is, the recommendation information number feature, and the second initial sample feature refers to the initial sample feature other than the first initial sample feature, that is, other features other than the account ID feature of the sample account and the recommendation information ID feature of the sample recommendation information, such as the gender feature of the sample account and the upload time feature of the sample recommendation information, etc. The first gated input sample feature refers to the first initial sample feature used as the gated input sample feature, the second gated input sample feature refers to the second initial sample feature used as the gated input sample feature, and the gated input screening information refers to the information for screening the second initial sample feature as the second gated input sample feature, for example, it may be a pre-set second gated input sample feature screening rule.

[0096] In this embodiment, the initial sample features may include a first initial sample feature and a second initial sample feature. When determining the gated input sample feature, all first initial sample features may be used as the first gated input sample feature, and for the second gated input sample feature, the second gated input sample feature may be performed according to the gated input screening information, thereby using the determined first gated input sample feature and the second gated input sample feature as the gated input sample feature. Since the first initial sample feature can be used to characterize and identify the sample account, or identify the sample recommendation information, the gated input sample feature screened in the above manner can, under the premise of improving the computational efficiency of the information recommendation model, realize that the output of the gated structure can be adaptively adjusted according to the identified sample account or sample recommendation information, thereby improving the flexibility of the output of the gated structure.

[0097] In the above embodiment, the server may use the first initial sample feature used to identify the sample account or sample recommendation information as the first gated input sample feature, and the second initial sample feature other than the first initial sample feature in the initial sample features may be screened based on the gated input screening information to obtain the second gated input sample feature, so that the output of the gated structure can be adaptively adjusted according to the identified sample account or sample recommendation information while improving the computational efficiency of the information recommendation model, thereby improving the flexibility of the output of the gated structure.

[0098] Figure 5 is a flow chart of an information recommendation method according to an exemplary embodiment. Figure 5 As shown, the information recommendation method is used in a server and includes the following steps.

[0099] In step S501, information recommendation features are obtained; wherein the information recommendation features are obtained based on account features of the target account and information features of the candidate recommended information.

[0100] The target account refers to the account to be recommended, and the candidate recommendation information refers to the recommendation information in the recommendation information library that can be used to recommend to the target account, such as videos contained in a video database. In this embodiment, when the server needs to recommend recommended information to an account, for example, when a user triggers a page refresh operation through his terminal, thereby triggering an information recommendation request for the server, the server can respond to the request, thereby obtaining the account features of the target account, and extracting the information features of each candidate recommendation information from the recommendation information library, and then, based on the obtained account features and information features, information recommendation features for input into the information recommendation model can be formed.

[0101] In step S502, the information recommendation features are input into the trained information recommendation model to obtain the predicted operation result of the target account for the candidate recommended information; wherein the information recommendation model is obtained by the training method of the information recommendation model described in any of the above embodiments.

[0102] The predicted operation result refers to the predicted operation status of the target account for each candidate recommended information output by the information recommendation model. For example, it can be the predicted operation probability of the target account for each candidate recommended information. In this embodiment, after the information recommendation model is trained by the information recommendation model training method provided by the above embodiment, the server can further input the information recommendation features obtained in step S501 into the trained information recommendation model, and then the predicted operation result of the target account for each candidate recommended information can be obtained through the output of the information recommendation model.

[0103] In step S503, according to the prediction operation result, target recommendation information for the target account is screened out from the candidate recommendation information, and the target recommendation information is recommended to the target account.

[0104] The target recommendation information refers to the recommendation information that is ultimately recommended to the target account among the candidate recommendation information. Finally, after obtaining the predicted operation result of the target account for each candidate recommendation information, the server can further filter out the target recommendation information for recommendation to the target account from each candidate recommendation information based on the predicted operation result. For example, the predicted operation result can be the predicted operation probability of the target account for each candidate recommendation information. The server can then recommend the candidate recommendation information with the highest predicted operation probability as the target recommendation information to the target account.

[0105] In the above information recommendation method, information recommendation features are obtained; wherein the information recommendation features are obtained based on the account features of the target account and the information features of the candidate recommended information; the information recommendation features are input into the trained information recommendation model to obtain the predicted operation results of the target account for the candidate recommended information; wherein the information recommendation model is obtained by the training method of the information recommendation model as described in any of the above embodiments; according to the predicted operation results, the target recommended information for the target account is screened out from the candidate recommended information, and the target recommended information is recommended to the target account. In the information recommendation method provided by the present disclosure, the trained information recommendation model can obtain the importance of the target sample features output by each feature extraction layer through the output value of the gating structure, and the prediction operation result output by the information recommendation model can be determined based on the target sample features and the output value of the gating structure, thereby realizing the training of the information recommendation model. Compared with the traditional technology of realizing information recommendation model training by fusing the features output by each feature extraction layer by directly adding them, the present disclosure can realize the fusion of features based on the importance of the features output by each feature extraction layer, and the above-mentioned importance can be determined by the initial sample features input to the information recommendation model, realizing the adaptive fusion of the features output by each feature extraction layer, thereby improving the recommendation accuracy of the trained information recommendation model.

[0106] In an exemplary embodiment, a recommendation method based on a multi-branch multi-depth network gating structure is also provided. This method is used in the scenario of recommendation network and can improve the prediction effect of the recommendation system. This method can enable the neural network model to have various features from shallow to deep, and allow each sample / user / item to adaptively fuse multi-depth branch features according to different importance levels through the gating mechanism, so as to obtain a more appropriate and accurate prediction value. Specifically, the following steps may be included:

[0107] First, we organize training samples based on historical data. Each sample consists of: user ID, i.e. user number, target item ID, i.e. target item number, user-side features (such as gender, browsing history, etc.), target item-side features (such as upload time, etc.), and some user-item context features (such as time and location of encounter, etc.). The original representation in each feature is generally a number or a string. For example, the feature of city has many feature value keys corresponding to different cities. We can use a finite set of numbers to represent different feature value keys (for example, key1 of the feature corresponds to city 1, and key2 corresponds to city 2). Each key of the feature will obtain a corresponding vector (within the same feature, different keys have the same dimension of the corresponding vector, but different values). For example, the key of the city feature = 'city 1' has its own unique vector, and the key of the city feature = 'city 2' has a vector with the same dimension as 'city 1' but different values. Therefore, by the above method, the key of this number or string is converted into a unique vector to realize feature vectorization.

[0108] Input sample, now take out the key vectors corresponding to these N features and concatenate them together to get input Input=[e1,e2,…e n ]. Input it into the neural network model. Finally, the estimated value pred = F (Input, W). Function F represents the network model, W represents the parameters in the network model, and pred is the estimated value of the user clicking / watching the video. This embodiment mainly acts on the above network model, and provides a multi-branch multi-depth network gate structure to make the estimated value of the neural network more accurate. The structure of the network model can be as follows Figure 2 shown.

[0109] Specifically, first extract the corresponding N vectors of the N features of the sample and merge them into a vector A = [e1, e2, ... e n There are many ways to merge, such as directly connecting all vectors in series, or performing other operations on local feature vectors (such as multiplication, neural network to generate new feature vectors) and then merging them together. This embodiment adopts the direct series connection method.

[0110] According to the merged vector A, the neural network is used to project it in turn (increasing the number of layers of the neural network) to obtain O1, O2, ..., O K There are K vectors output in total. That is:

[0111] O1=F1(A*W1)

[0112] O k =F k (O k-1 *W k )

[0113] The function F represents a single-layer feature extraction layer of the neural network, F(X,W) = σ(X*W), σ represents the ReLU activation function, that is, the linear rectification function, and W is a parameter in the neural network, where W k represents the feature extraction layer parameters of the kth feature extraction layer. In this way, vector outputs of K depths are obtained, which represent the information extracted by the model at different depths.

[0114] Take vector A as the input of the gate (as long as the input contains the vector of user ID features or the vector of item ID features, so that user-level / item-level adaptive branch gating can be achieved), and obtain O1, O2, …, O through the gating mechanism. K The gate value of each branch [g1,g2,…,g K ] (K values), this gating value indicates the importance of the corresponding branch output, or the fusion coefficient. This gating mechanism uses the MLP neural network structure: [g1,g2,…,g k ]=F g (A,W g )=σ(A*W g ), where σ is the sigmoid activation function, that is, the S-shaped double bending function, and W g It represents the structural parameters corresponding to the MLP neural network structure used for the gating structure.

[0115] The output of multiple branches [O1, O2, ..., O K ] respectively correspond to multiplying the gate value [g1,g2,…,g K ], adding them together to form a new vector Among them, k represents the kth target sample feature, and g k It represents the output value of the gate parameter corresponding to the kth target sample feature. This vector is the new feature of sample personalization fusion multi-branch multi-depth. The final estimated value can be obtained by passing this vector through the final MLP neural network:

[0116] pred=F(O * *W * )=σ(O * *W)

[0117] Where σ is the sigmoid activation function, that is, the S-shaped double bending function, and W represents the model parameters of the final MLP neural network, and the obtained pred is the user's estimated click value for the video.

[0118] In the above embodiment, the requirements for deep and shallow features in the recommendation system can be met, and the gating mechanism can allow samples / users / items to adaptively find the best feature fusion method to obtain more accurate estimates. This can effectively improve the effect of the recommendation model and obtain business benefits. In addition, the architecture of the model greatly improves the potential of the entire model. By providing a concise, efficient and decoupled architecture, it is convenient to integrate a variety of other effective model structures into this structure to further improve the model effect.

[0119] It should be understood that, although the various steps in the flowchart of the present disclosure are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the figure may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of the steps or stages in other steps.

[0120] It can be understood that the same / similar parts between the various embodiments of the above method in this specification can refer to each other, and each embodiment focuses on the differences from other embodiments. For related points, please refer to the description of other method embodiments.

[0121] Figure 6 is a block diagram of a training device for an information recommendation model according to an exemplary embodiment. Figure 6 The device includes a sample data acquisition unit 601, a prediction result acquisition unit 602 and a recommendation model training unit 603.

[0122] The sample data acquisition unit 601 is configured to acquire initial sample features and actual operation results of the sample account for the sample recommendation information; the initial sample features are acquired according to the sample account features of the sample account and the sample information features of the sample recommendation information;

[0123] The prediction result acquisition unit 602 is configured to input the initial sample features into the information recommendation model to be trained, and obtain the prediction operation result of the sample account for the sample recommendation information; the information recommendation model includes a gating structure and multiple feature extraction layers, the multiple feature extraction layers are used to output multiple target sample features for the initial sample features, the output value of the gating structure is obtained according to the initial sample features, and is used to characterize the importance of each target sample feature, and the prediction operation result is obtained according to the multiple target sample features and the output value of the gating structure;

[0124] The recommendation model training unit 603 is configured to train the information recommendation model to be trained based on the difference between the predicted operation result and the actual operation result to obtain a trained information recommendation model.

[0125] In an exemplary embodiment, the prediction result acquisition unit 602 is further configured to execute the input of the initial sample features into the information recommendation model to be trained, and obtain multiple target sample features for the initial sample features through multiple feature extraction layers of the information recommendation model; input the initial sample features into the gating structure, and obtain the output value of the gating structure corresponding to each target sample feature through the gating structure; perform weighted processing on each target sample feature according to the output value of the gating structure, and obtain the prediction operation result using the result of the weighted processing.

[0126] In an exemplary embodiment, the prediction result acquisition unit 602 is further configured to execute acquiring the current feature extraction layer from multiple feature extraction layers, and acquiring the current input sample features for inputting the current feature extraction layer; inputting the current input sample features into the current feature extraction layer, and obtaining the target sample features corresponding to the current feature extraction layer through the feature extraction layer parameters corresponding to the current feature extraction layer.

[0127] In an exemplary embodiment, the prediction result acquisition unit 602 is further configured to use the initial sample features as the current input sample features when the current feature extraction layer is the first feature extraction layer of the information recommendation model to be trained; and to use the target sample features corresponding to the previous feature extraction layer as the current input sample features when the current feature extraction layer is not the first feature extraction layer of the information recommendation model to be trained.

[0128] In an exemplary embodiment, the prediction result acquisition unit 602 is further configured to acquire gated input sample features from the initial sample features; input the gated input sample features into the gated structure, and obtain the output value of the gated structure through the gated structure.

[0129] In an exemplary embodiment, the initial sample feature is composed of a first initial sample feature and a second initial sample feature; the first initial sample feature is an initial sample feature used to identify a sample account or to identify sample recommendation information, and the second initial sample feature is an initial sample feature other than the first initial sample feature among the initial sample features; the prediction result acquisition unit 602 is further configured to perform a filter according to preset gated input filtering information, and filter out a second gated input sample feature from the second initial sample feature; use the first initial sample feature as the first gated input sample feature, and use the first gated input sample feature and the second gated input sample feature as gated input sample features.

[0130] Figure 7 is a block diagram of an information recommendation device according to an exemplary embodiment. Figure 7 The device includes a recommendation feature acquisition unit 701, a model prediction acquisition unit 702 and a target information recommendation unit 703.

[0131] The recommendation feature acquisition unit 701 is configured to acquire information recommendation features; wherein the information recommendation features are acquired according to the account features of the target account and the information features of the candidate recommended information;

[0132] The model prediction acquisition unit 702 is configured to input the information recommendation feature into the trained information recommendation model to obtain the prediction operation result of the target account for the candidate recommended information; wherein the information recommendation model is obtained by the information recommendation model training method as described in any of the above embodiments;

[0133] The target information recommendation unit 703 is configured to perform, according to the prediction operation result, screening out target recommendation information for the target account from the candidate recommendation information, and recommending the target recommendation information to the target account.

[0134] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0135] Figure 8 1 is a block diagram of an electronic device 800 for training an information recommendation model or for information recommendation according to an exemplary embodiment. For example, the electronic device 800 may be a server. Figure 8 , the electronic device 800 includes a processing component 820, which further includes one or more processors, and a memory resource represented by a memory 822 for storing instructions executable by the processing component 820, such as an application. The application stored in the memory 822 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 820 is configured to execute the instructions to perform the above method.

[0136] The electronic device 800 may further include a power supply component 824 configured to perform power management of the electronic device 800, a wired or wireless network interface 826 configured to connect the electronic device 800 to a network, and an input / output (I / O) interface 828. The electronic device 800 may operate based on an operating system stored in the memory 822, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, or the like.

[0137] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory 822 including instructions, and the above instructions can be executed by a processor of the electronic device 800 to perform the above method. The storage medium can be a computer-readable storage medium, for example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0138] In an exemplary embodiment, a computer program product is further provided. The computer program product includes instructions. The instructions can be executed by a processor of the electronic device 800 to complete the above method.

[0139] It should be noted that the above-mentioned devices, electronic devices, computer-readable storage media, computer program products, etc. may also include other implementation methods according to the description of the method embodiments. The specific implementation methods can refer to the description of the relevant method embodiments, which will not be described one by one here.

[0140] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.

[0141] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A training method for an information recommendation model, characterized in that: include: Obtaining initial sample characteristics and actual operation results of sample accounts for sample recommendation information; The initial sample characteristics include sample account characteristics of the sample account and sample information characteristics of the sample recommendation information; The sample recommendation information includes a sample video to be recommended, the sample information feature includes a video feature of the sample video to be recommended, and the actual operation result is used to indicate whether the sample account performs a user interaction operation on the sample video to be recommended; The initial sample features are input into the information recommendation model to be trained to obtain the predicted operation results of the sample account for the sample recommendation information; the information recommendation model includes a gating structure and multiple feature extraction layers, the multiple feature extraction layers are used to output multiple target sample features for the initial sample features, the output value of the gating structure is obtained according to the initial sample features, and is used to characterize the importance of each target sample feature, and the predicted operation results are obtained according to the multiple target sample features and the output value of the gating structure; the multiple feature extraction layers correspond to different feature extraction layers of the information recommendation model, and each target sample feature is a feature extracted by the information recommendation model at different depths; the predicted operation result is used to characterize the predicted user interaction operation probability of the sample account for the sample video to be recommended; Based on the difference between the predicted operation result and the actual operation result, the information recommendation model to be trained is trained to obtain a trained information recommendation model.

2. The method according to claim 1, characterized in that The step of inputting the initial sample features into the information recommendation model to be trained to obtain the prediction operation result of the sample account for the sample recommendation information includes: Inputting the initial sample features into the information recommendation model to be trained, and obtaining multiple target sample features for the initial sample features through multiple feature extraction layers of the information recommendation model; Inputting the initial sample features into the gating structure, and obtaining the output values ​​of the gating structure corresponding to each target sample feature through the gating structure; The characteristics of each target sample are weighted according to the output value of the gating structure, and the prediction operation result is obtained using the result of the weighted processing.

3. The method according to claim 2, characterized in that The multiple feature extraction layers of the recommendation model are used to obtain multiple target sample features for the initial sample features, including: From the plurality of feature extraction layers, obtain a current feature extraction layer, and obtain current input sample features for inputting the current feature extraction layer; The current input sample feature is input into the current feature extraction layer, and the target sample feature corresponding to the current feature extraction layer is obtained through the feature extraction layer parameters corresponding to the current feature extraction layer.

4. The method according to claim 3, characterized in that The step of obtaining the current input sample features for inputting into the current feature extraction layer includes: In a case where the current feature extraction layer is the first feature extraction layer of the information recommendation model to be trained, using the initial sample feature as the current input sample feature; In the case that the current feature extraction layer is not the first feature extraction layer of the information recommendation model to be trained, the target sample feature corresponding to the previous feature extraction layer is used as the current input sample feature.

5. The method according to claim 2, characterized in that: The step of inputting the initial sample feature into the gating structure and obtaining the output value of the gating structure corresponding to each target sample feature through the gating structure includes: Obtaining gated input sample features from the initial sample features; The gated input sample feature is input into the gated structure, and the output value of the gated structure is obtained through the gated structure.

6. The method according to claim 5, characterized in that The initial sample feature is composed of a first initial sample feature and a second initial sample feature; the first initial sample feature is an initial sample feature used to identify the sample account or to identify the sample recommendation information, and the second initial sample feature is an initial sample feature among the initial sample features except the first initial sample feature; The step of obtaining gated input sample features from the initial sample features includes: According to the preset gated input screening information, a second gated input sample feature is screened out from the second initial sample feature; The first initial sample feature is used as a first gated input sample feature, and the first gated input sample feature and the second gated input sample feature are used as the gated input sample feature.

7. An information recommendation method, characterized in that: include: Acquire information recommendation features; wherein the information recommendation features include account features of the target account and information features of candidate recommendation information; the candidate recommendation information includes candidate videos to be recommended, and the information features include video features of the candidate videos to be recommended; Input the information recommendation feature into the trained information recommendation model to obtain the predicted operation result of the target account for the candidate recommended information; wherein the information recommendation model is obtained by the training method of the information recommendation model according to any one of claims 1 to 6; the predicted operation result is used to characterize the predicted user interaction operation probability of the target account for the candidate video to be recommended; According to the prediction operation result, target recommendation information for the target account is screened out from the candidate recommendation information, and the target recommendation information is recommended to the target account.

8. A training device for an information recommendation model, characterized in that: include: A sample data acquisition unit, configured to acquire initial sample features and actual operation results of the sample account for sample recommendation information; The initial sample characteristics include sample account characteristics of the sample account and sample information characteristics of the sample recommendation information; The sample recommendation information includes a sample video to be recommended, the sample information feature includes a video feature of the sample video to be recommended, and the actual operation result is used to indicate whether the sample account performs a user interaction operation on the sample video to be recommended; A prediction result acquisition unit is configured to execute inputting the initial sample features into the information recommendation model to be trained, and obtain the prediction operation result of the sample account for the sample recommendation information; the information recommendation model includes a gating structure and multiple feature extraction layers, the multiple feature extraction layers are used to output multiple target sample features for the initial sample features, the output value of the gating structure is obtained according to the initial sample features, and is used to characterize the importance of each target sample feature, and the prediction operation result is obtained according to the multiple target sample features and the output value of the gating structure; the multiple feature extraction layers correspond to different feature extraction layers of the information recommendation model, and each of the target sample features is a feature extracted by the information recommendation model at different depths; the prediction operation result is used to characterize the predicted user interaction operation probability of the sample account for the sample video to be recommended; The recommendation model training unit is configured to perform training on the information recommendation model to be trained based on the difference between the predicted operation result and the actual operation result to obtain a trained information recommendation model.

9. The device according to claim 8, characterized in that The prediction result acquisition unit is further configured to input the initial sample feature into the information recommendation model to be trained, and obtain multiple target sample features for the initial sample feature through multiple feature extraction layers of the information recommendation model; input the initial sample feature into the gating structure, and obtain the output value of the gating structure corresponding to each target sample feature through the gating structure; The characteristics of each target sample are weighted according to the output value of the gating structure, and the prediction operation result is obtained using the result of the weighted processing.

10. The device according to claim 9, characterized in that The prediction result acquisition unit is further configured to execute acquiring a current feature extraction layer from the multiple feature extraction layers, and acquiring current input sample features for inputting the current feature extraction layer; inputting the current input sample features into the current feature extraction layer, and obtaining target sample features corresponding to the current feature extraction layer through feature extraction layer parameters corresponding to the current feature extraction layer.

11. The device according to claim 10, characterized in that The prediction result acquisition unit is further configured to use the initial sample feature as the current input sample feature when the current feature extraction layer is the first feature extraction layer of the information recommendation model to be trained; In the case that the current feature extraction layer is not the first feature extraction layer of the information recommendation model to be trained, the target sample feature corresponding to the previous feature extraction layer is used as the current input sample feature.

12. The device according to claim 9, characterized in that The prediction result acquisition unit is further configured to acquire a gated input sample feature from the initial sample feature; input the gated input sample feature into the gate structure, and obtain an output value of the gate structure through the gate structure.

13. The device according to claim 12, characterized in that The initial sample feature is composed of a first initial sample feature and a second initial sample feature; the first initial sample feature is an initial sample feature used to identify the sample account, or used to identify the sample recommendation information, and the second initial sample feature is an initial sample feature other than the first initial sample feature among the initial sample features; the prediction result acquisition unit is further configured to perform a filter according to preset gated input filtering information, and filter out a second gated input sample feature from the second initial sample feature; use the first initial sample feature as the first gated input sample feature, and use the first gated input sample feature and the second gated input sample feature as the gated input sample features.

14. An information recommendation device, characterized in that: include: A recommendation feature acquisition unit is configured to acquire information recommendation features; wherein the information recommendation features include account features of a target account and information features of candidate recommendation information; the candidate recommendation information includes candidate videos to be recommended, and the information features include video features of the candidate videos to be recommended; A model prediction acquisition unit, configured to input the information recommendation feature into a trained information recommendation model to obtain a predicted operation result of the target account for the candidate recommended information; wherein the information recommendation model is obtained by the training method of the information recommendation model according to any one of claims 1 to 6; the predicted operation result is used to characterize the predicted user interaction operation probability of the target account for the candidate video to be recommended; The target information recommendation unit is configured to perform, according to the prediction operation result, filtering out target recommendation information for the target account from the candidate recommendation information, and recommending the target recommendation information to the target account.

15. A server, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the training method of the information recommendation model as described in any one of claims 1 to 6, or the information recommendation method as described in claim 7.

16. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by a processor of a server, the server is enabled to execute the training method of the information recommendation model as described in any one of claims 1 to 6, or the information recommendation method as described in claim 7.

17. A computer program product, comprising instructions, characterized in that: When the instruction is executed by the processor of the server, the server is enabled to execute the training method of the information recommendation model as described in any one of claims 1 to 6, or the information recommendation method as described in claim 7.

Images